Excellence is a habit: Enhancing predictions of language impairment by identifying stable features in clinical perfusion scans

Perfusion images guide acute stroke management, yet few studies have been able to systematically investigate CT perfusion collected during routine care because the measures are stored in proprietary formats incompatible with conventional research analysis pipelines. We illustrate the potential of harnessing granular data from these routine scans by using them to identify the association between specific areas of hypoperfusion and severity of object naming impairment in 43 acute stroke patients. Traditionally, similar analyses in such sample sizes face a dilemma--simple models risk being too constrained to make accurate predictions, while complex models risk overfitting and producing poor out-of-sample predictions. We demonstrate that evaluating the stability rather than out-of-sample predictive capacity of features in a nested cross-validation scheme can be an effective way of controlling model complexity and stabilizing model estimates across a variety of different regression techniques. Specifically, we show that introducing this step can determine model significance, even when the regression model already contains an embedded feature selection or dimensionality reduction step, or if a subset of features is manually selected prior to training based on expert knowledge. After improving model performance using more complex regression techniques, we discover that object naming performance relies on an extended language network encompassing regions thought to play a larger role in different naming tasks, right hemisphere regions distal to the site of injury, and regions and tracts that are less typically associated with language function. Our findings especially emphasize the role of the left superior temporal gyrus, uncinate fasciculus, and posterior insula in successful prediction of object naming impairment. Collectively, these results highlight the untapped potential of clinical CT perfusion images and demonstrate a flexible framework for enabling prediction in the limited sample sizes that currently dominate clinical neuroimaging.